> On Oct 31, 2018, at 5:22 PM, David Rowley <david.row...@2ndquadrant.com> > wrote: > > On 1 November 2018 at 04:40, John Naylor <jcnay...@gmail.com> wrote: >> Thanks for doing this. I haven't looked at the rendered output yet, >> but I have some comments on the content. >> >> + <entry>Maximum Relation Size</entry> >> + <entry>32 TB</entry> >> + <entry>Limited by 2^32 pages per relation</entry> >> >> I prefer "limited to" or "limited by the max number of pages per >> relation, ...". I think pedantically it's 2^32 - 1, since that value >> is used for InvalidBlockNumber. More importantly, that seems to be for >> 8kB pages. I imagine this would go up with a larger page size. Page >> size might also be worth mentioning separately. Also max number of >> relation file segments, if any. > > Thanks for looking at this. > > I've changed this and added mention of BLKSIZE. I was a bit unclear > on how much internal detail should go into this.
It’s a bit misleading to say “Can be increased by increasing BLKSZ and recompiling”, since you’d also need to re initdb. Given that messing with BLKSZ is pretty uncommon I would simply put a note somewhere that mentions that these values assume the default BLKSZ of 8192. >> + <entry>Maximum Columns per Table</entry> >> + <entry>250 - 1600</entry> >> + <entry>Depending on column types. (More details here)</entry> >> >> Would this also depend on page size? Also, I'd put this entry before this >> one: >> >> + <entry>Maximum Row Size</entry> >> + <entry>1600 GB</entry> >> + <entry>Assuming 1600 columns, each 1 GB in size</entry> >> >> A toast pointer is 18 bytes, according to the docs, so I would guess >> the number of toasted columns would actually be much less? I'll test >> this on my machine sometime (not 1600GB, but the max number of toasted >> columns per tuple). > > I did try a table with 1600 text columns then inserted values of > several kB each. Trying with BIGINT columns the row was too large for > the page. I've never really gotten a chance to explore these limits > before, so I guess this is about the time. Hmm… 18 bytes doesn’t sound right, at least not for the Datum. Offhand I’d expect it to be the small (1 byte) varlena header + an OID (4 bytes). Even then I don’t understand how 1600 text columns would work; the data area of a tuple should be limited to ~2000 bytes, and 2000/5 = 400.